When Does Content-Aware Freshness Pay? Visual Semantic Age of Information for Multi-Camera Uplink Scheduling
Abstract — Age of Information grows linearly with time and is blind to content, so a camera observing an empty road is treated as urgently as one observing dense traffic. We introduce Visual Semantic Age of Information (V-SAoI), in which the per-slot age increment is scaled by scene dynamics measured causally at the sensor, requiring no receiver feedback and no source model. A Lyapunov drift-plus-penalty analysis yields a per-slot index rule of complexity O(K log K) under long-term average power constraints. We evaluate on synthetic traces and on a mixed real benchmark drawn from KITTI and MOT17, judging every policy on downstream perception fidelity rather than on the age metric being optimised, and bounding the achievable opportunity with an oracle. Content-aware scheduling reduces perception loss by 6.61% on synthetic traces but only 2.46% on real tracks, where it captures 84.5% of the available headroom. We show the limiting factor is the headroom itself, which barely responds to increased camera heterogeneity, and we identify the conditions under which content-aware freshness is worth deploying.
Keywords — Age Of Information; Semantic Communications; Multi-Camera Uplink; Lyapunov Optimization; Resource Allocation; Edge Vision
I. Introduction
Age of Information (AoI) has become the standard measure of status freshness in wireless monitoring systems [1]. It grows linearly with the time elapsed since the last delivered update, and its underlying assumption is that the value of an update depends on when it arrived rather than on what it contains. For visual sensing this assumption is hard to defend. A roadside camera watching an empty carriageway conveys almost nothing new after fifty idle slots. A camera watching dense turning traffic may already be badly stale after five. Linear AoI scores these two situations identically, and any scheduler driven by it will allocate uplink resources accordingly.
Several lines of work relax the assumption. Urgency of Information introduces a context-dependent weight together with a nonlinear cost [2], while Age of Incorrect Information couples freshness to the estimation error observed at the receiver [3]. More recent studies address semantics directly for video transmission [4] and for latent feature freshness in multi-camera perception [5]. In each case the quantity that modulates the age is either supplied externally, derived from an abstract Markov source model, or recovered through receiver-side state estimation and feedback [6].
We start from a simpler observation. For visual sources this quantity does not need to be modelled or fed back, because it can be measured at the sensor itself. The rate at which a scene changes is visible in the motion field between consecutive frames, which a camera already computes. Block motion vectors are produced as a by-product of H.264 and HEVC encoding, so the signal is available at negligible additional cost. This yields a causal, source-local urgency measure that requires no receiver feedback and no channel knowledge beyond what scheduling already needs.
We define Visual Semantic Age of Information (V-SAoI), in which the per-slot age increment is scaled by the measured visual dynamics of the scene. A Lyapunov drift-plus-penalty analysis then yields a closed-form per-slot index rule of complexity O(K log K) under per-camera long-term average power constraints. The contributions are threefold. First, we introduce a content-scaled freshness metric whose driving signal is measured causally at the source and establish it as a conservative surrogate for latent feature drift. Second, we derive the corresponding scheduler and its index rule. Third, we evaluate it in a deliberately non-circular way: every policy is judged on a downstream perception metric rather than on the age metric being optimised, and a myopic oracle bounds the headroom available to any content-aware policy.
The evaluation leads to a result worth stating at the outset. Content-aware scheduling is not uniformly beneficial. Across synthetic traces and two real benchmarks we find that the achievable gain is governed primarily by how much the cameras differ from one another in scene dynamics, and that this heterogeneity can be measured before any scheduler is deployed. We therefore report both a regime in which the gain is substantial and one in which it is small, and we identify the condition that separates them. Establishing when a mechanism pays seems more useful than reporting a single headline figure.
II. System Model
A. Network and Channel
Consider K edge cameras uploading to a common edge server over M < K orthogonal uplink resources in slotted time t = 1, 2, …. Let xk(t) ∈ {0, 1} indicate that camera k is granted a resource in slot t, subject to a per-slot budget of M grants. The channel gain |hk(t)|2 comprises a fixed large-scale component and independent Rayleigh small-scale fading. Delivering a payload of spectral efficiency D over bandwidth B requires the power given in (1). Camera k is feasible in slot t if Pkreq(t) does not exceed a peak constraint Pmax, and is otherwise in outage. Each camera is further subject to a long-term average power constraint Pavg.
B. Measuring Visual Dynamics
Each camera computes a scalar γk(t) ≥ 0 from consecutive frames, measuring the rate at which its scene is changing. Dense optical flow magnitude, codec motion vector magnitude, or the mean displacement of tracked object centroids are all admissible. We adopt the last of these for evaluation, since it is directly recoverable from annotated benchmark traces and therefore avoids conflating the metric with the accuracy of a particular flow estimator. Coordinates are normalised by frame dimensions so that γk(t) is comparable across cameras of differing resolution.
Traces captured at different native frame rates are resampled onto a common slot grid by nearest-frame selection before γ is computed, rather than being rescaled afterwards. This step matters more than it might appear. A sequence captured at 30 Hz exhibits per-frame displacement roughly three times smaller than the same scene at 10 Hz, so pooling heterogeneous sources without resampling manufactures apparent heterogeneity that is an artifact of sampling rather than a property of the scene.
The urgency increment follows from γ through the clipped-linear map of (2), which satisfies 1 ≤ g(·) ≤ 1 + β. Three properties are required. Linearity below γmax supports the bound in Lemma 1. Boundedness above supplies the second-moment condition needed by the drift-plus-penalty analysis in Section III. The unit floor guarantees that a permanently static camera still accumulates age and is eventually served, which rules out starvation. The constant γmax is a calibration parameter, set from the observed motion distribution of the deployment in the same way a fielded system would set it.
One caveat applies to camera-mounted platforms, where γ conflates object motion with ego-motion. Decomposing the per-frame displacement field into a coherent component and the residual about it, we estimate that ego-motion contributes roughly 43% of γ on driving sequences, against under 3% on verified fixed cameras. We regard this as legitimate rather than as contamination, since a rapidly moving camera genuinely does go stale faster, but it does mean that on driving-only data the measured motion intensity partly reflects platform speed.
C. Visual Semantic Age of Information
V-SAoI evolves according to (3), and reduces to conventional AoI when β = 0. Because g depends only on frames already captured at camera k, the update is causal and entirely source-local. Reporting γk(t) to the scheduler costs a few bits per camera per slot, and Section IV shows that two bits are sufficient.
Lemma 1 (Conservative surrogate): Let zk(t) denote a latent representation of frame t that is L-Lipschitz in the scene motion field, and let Uk(t) be the last delivery instant. Then (4) holds with C = L. Moreover, for γ ≤ γmax the accumulated increment obeys (5).
Proof: The first inequality is the triangle inequality applied along the frame sequence. The second follows from the Lipschitz assumption. The third follows from the linearity of g below the clipping threshold. □
V-SAoI therefore upper-bounds the true semantic staleness and remains computable at the source without acknowledgement-based receiver state. The bound is loose under oscillatory motion, where accumulated displacement exceeds net displacement. We treat this as deliberate conservatism, since such scenes are genuinely ambiguous when judged from motion magnitude alone.
D. Problem Formulation
The objective is to minimise the long-term time-averaged weighted V-SAoI as stated in (6), over the scheduling decisions {xk(t)}, subject to the per-slot grant budget, the feasibility condition of Section II-A, and the long-term average power constraint for every camera. The result is a stochastic combinatorial program in which slots are coupled through the power constraints.
(1) Pkreq(t) = (2D/B − 1) σ2 / |hk(t)|2
(2) g(γ) = 1 + β · min(γ, γmax) / γmax
(3) Δksem(t+1) = [ Δksem(t) + g(γk(t)) ] · [ 1 − xk(t) ]
(4) ‖zk(t) − zk(Uk(t))‖ ≤ Στ ‖zk(τ) − zk(τ−1)‖ ≤ C · Στ γk(τ), τ = Uk(t)+1, …, t
(5) Στ g(γk(τ)) ≥ (β / γmax) · Στ γk(τ)
(6) min{xk(t)} lim supT→∞ (1/T) Σt=1T Σk=1K wk Δksem(t)
III. Drift Plus Penalty Scheduler
To handle the long-term power constraint we introduce a virtual queue for each camera as in (7), and define the Lyapunov function over the joint state of ages and queues as in (8).
Because the scheduling variable is binary, squaring the recursion in (3) gives a factor of 1 − xk rather than its square, and the age contribution to the one-slot drift separates as shown in (9), where the first term does not depend on the decision. The penalty must be evaluated at slot t + 1. Evaluating it at t, as is sometimes done, leaves it already determined at the moment x(t) is chosen, so it contributes nothing to the per-slot problem.
Minimising the resulting bound reduces to maximising a weighted sum of grants, with the index of (10). The three terms admit a direct reading. The first grows quadratically in the post-increment semantic age, so a camera whose scene is changing rapidly climbs the ranking faster than one whose scene is static. The second is the linear penalty on the objective. The third prices the transmission by the power it would consume against the camera's accumulated energy debt, so a camera in deep fade is deferred until either its channel improves or its age makes the cost worth paying.
The scheduler grants the M feasible cameras with the largest index, at cost O(K log K) per slot. Since g is bounded above and the required power has bounded second moment under the peak constraint, the standard order 1/V and order V trade-off between objective and constraint violation applies [7].
Remark (tie structure): When β = 0 every camera advances by the same increment, and the index ties at the grant boundary in a large fraction of slots. In our setting this occurs in 39.4% of slots, so allocation is decided by the tie-breaking rule rather than by the metric itself. Content scaling reduces the tie rate to 0.3%. V-SAoI therefore induces a near-total order on urgency where linear AoI induces only a coarse one, which is a secondary benefit of the metric independent of the scheduling gain. All policies in Section IV use an index-independent random tie-break so that no policy benefits from the ordering of camera indices.
(7) Qk(t+1) = max[ Qk(t) + xk(t) Pkreq(t) − Pavg , 0 ]
(8) L(Θ(t)) = (1/2) Σk=1K [ α (Δksem(t))2 + Qk2(t) ]
(9) (α/2)[ (Δksem + gk)2 − (Δksem)2 ] − (α/2) xk (Δksem + gk)2
(10) Wk(t) = (α/2)[Δksem(t) + g(γk(t))]2 + V wk [Δksem(t) + g(γk(t))] − Qk(t) Pkreq(t)
IV. Performance Evaluation
A. Simulation Setup
We consider K = 12 edge cameras sharing M = 3 uplink resources over a Rayleigh fading channel with independent large-scale gains. Results are averaged over 48 independent realisations, with the first fifth of each run discarded as transient. The urgency parameter is β = 4 and the calibration constant γmax is set to the 88th percentile of the pooled dynamics distribution of each trace set.
Three trace sets are used. The synthetic set drives twelve cameras with Ornstein-Uhlenbeck object motion spanning a 16-fold range of mean scene dynamics, and serves as a controlled study in which heterogeneity can be dialled directly. The KITTI-only set uses twelve driving sequences from the KITTI tracking benchmark [8] and spans 4.48-fold. The mixed set combines five KITTI driving scenes with four moving pedestrian scenes and three fixed surveillance cameras from MOT17 [9] , spanning 7.62-fold. The mixed set is the primary real benchmark, since twelve sequences captured by a single vehicle-mounted rig do not instantiate a model of K cameras observing distinct scenes. All sequences are resampled to a common 10 Hz grid before γ is computed, for the reason given in Section II-B.
B. Baselines and Evaluation Metric
Six policies are compared. The proposed V-SAoI scheduler uses the index of (10). The linear AoI scheduler uses the identical drift-plus-penalty machinery with g ≡ 1, so the difference between the two isolates the effect of content scaling and nothing else. We further compare a max AoI times channel rule, a max throughput rule, round robin, and a myopic oracle that ranks cameras by their true instantaneous staleness. All six operate under the same per-camera average power constraint, enforced by an identical energy token bucket, so no policy is favoured by a differing energy allowance.
Policies are evaluated on downstream perception fidelity rather than on the age metric being optimised. The metric is the mean intersection over union between the bounding boxes held at the server from the last delivered frame and the ground truth boxes at the current instant, matched by track identity, with objects present now but absent from the stale set scoring zero. We report the perception loss, defined as one minus this quantity. The myopic oracle serves to bound the opportunity available to any content-aware policy, which proves to be the decisive quantity.
C. Achievable Headroom
Table 1 reports the oracle headroom, that is, the loss reduction achieved by the myopic oracle over linear AoI, which is the ceiling for any content-aware scheduler on the given data. The headroom on real tracks peaks near 3% and is largely insensitive to camera heterogeneity. Increasing the spread of scene dynamics from 4.48-fold to 7.62-fold raised it from 2.67% to 2.91% at the nominal budget, an increase of under a quarter of a percentage point for a 1.7-fold increase in heterogeneity. On synthetic traces the corresponding headroom reaches 7.5%.
This reframes the problem. The limiting quantity is not the scheduler but the opportunity, and the opportunity is a property of the perception task on real footage.
D. Perception Loss and Scheduling Gain
Table 2 summarises the main results across the three trace sets at the nominal and high power budgets. Figure 1 plots perception loss against the power budget for all six policies on the synthetic set, where the separation between policies is visible. V-SAoI tracks the oracle closely across the range and overtakes it at the highest budget, for the reason developed in Section IV-E. On real tracks the six curves compress into a band of roughly three percentage points, so we summarise them in Table 2 rather than plotting them.
Figure 2 places gain and headroom side by side for the three trace sets against their measured γ spread. The gap between the two curves is small throughout, which is the central observation of this work: V-SAoI captures most of what is available in every regime tested, and the variation across regimes is driven by the headroom rather than by the scheduler.
The mechanism operates as intended on real data. Figure 3 shows the share of uplink grants against each camera's measured scene dynamics on the mixed set. Under V-SAoI the correlation is 0.972, with grants ranging from 6.31% to 10.05%. Under linear AoI the correlation is 0.056 and the spread is 0.29 percentage points, that is, effectively flat. The scheduler identifies and prioritises the dynamic cameras. It simply gains little by doing so.
E. Starvation Behaviour of the Myopic Oracle
The oracle degrades as the power budget grows and falls below linear AoI at the highest setting, as Table 1 shows. Its grant concentration, measured as the ratio of largest to smallest camera share, rises from 3.2 to 7.9 to 10.2 across the budget range, reaching 20.17% for one sequence against 1.98% for another.
The cause is starvation. Greedy true-staleness repeatedly serves cameras whose staleness rebounds immediately after each delivery, owing to heavy object churn, while neglecting cameras that could be held near-perfect at low cost. The quadratic age term in (10) enforces a service floor that instantaneous staleness does not, which is why V-SAoI exceeds the oracle above the mid-range budget. It follows that the fraction of headroom captured is meaningful only at moderate budgets, since the denominator vanishes and then changes sign.
F. Ablation
Table 3 decomposes the gain on the mixed set at Pavg = 0.9. Camera heterogeneity is removed by equalising the mean dynamics across cameras. Temporal variation is removed by replacing each camera's dynamics signal with its own time average while leaving the underlying scene intact. The two effects contribute almost equally and together account for essentially the entire gain. This differs from the synthetic study, where heterogeneity dominated by roughly three to one. Real footage contains bursts, occlusions and stop-start behaviour that a stationary motion model understates, and the scheduler exploits them.
G. Robustness of the Calibration Constant
Sweeping γmax across the 50th to 99th percentile of the pooled dynamics distribution changes the gain only between 2.15% and 2.46%. The calibrated setting is the maximum but exceeds the median setting by 0.13 percentage points, so performance does not depend on precise calibration. Placing γmax far outside the data range, so that the urgency map is effectively constant, reduces the gain to 0.93%, confirming that content scaling rather than the particular constant is responsible.
On the synthetic study, quantising the reported dynamics signal to two bits per camera per slot recovers the full gain, and one bit recovers 66% of it. The uplink signalling cost of V-SAoI is therefore negligible relative to any payload.
| Conditions | Gain (%) | Change (pp) |
|---|---|---|
| Full (heterogenous, time varying) | 2.46 | |
| Homogenous Cameras | 1.31 | -1.55 |
| Time-averaged γ signal | 1.26 | -1.20 |
| Both removed | 0.19 | -2.27 |
| Trace set | γ spread | Gain at P = 0.9 (%) | Gain at P = 3.5 (%) | Headroom captured (%) |
|---|---|---|---|---|
| Synthetic | 16x | 6.61 | 10.66 | 87 |
| KITTI-only | 4.48x | 1.82 | 2.26 | 68 |
| Mixed | 7.62x | 2.46 | 3.25 | 84.6 |
| P_avg | KTTI-Only (%) | Mixed (%) |
|---|---|---|
| 0.60 | 1.54 | 1.51 |
| 0.90 | 2.67 | 2.91 |
| 1.40 | 2.67 | 2.48 |
| 2.20 | 1.74 | 0.98 |
| 3.50 | 0.76 | -0.88 |
V. Coclusion
We introduced V-SAoI, a freshness metric whose age increment is scaled by scene dynamics measured causally at the camera, together with a drift-plus-penalty scheduler admitting a closed-form per-slot index rule of complexity O(K log K). Evaluated against downstream perception fidelity rather than against itself, the method reduces perception loss by 6.61% on synthetic traces and by 2.46% on a mixed real benchmark drawn from KITTI and MOT17.
The bounding result is more useful than the headline. On real tracks the headroom available to any content-aware scheduler is close to 3%, and it responds only weakly to increased heterogeneity of the camera set. V-SAoI captures 84.5% of that headroom, so the limitation lies in the opportunity rather than in the mechanism. Because the headroom can be measured before deployment using the oracle construction given here, an operator can determine in advance whether content-aware freshness justifies its complexity for a given camera set. We also observed that greedy true-staleness scheduling becomes actively harmful at high power, since it starves cameras that could cheaply be held fresh, and that a quadratic age term corrects this.
Two limitations bear stating. The dynamics signal is computed from ground truth annotations, which idealises what a deployed camera would obtain from optical flow or codec motion vectors, although the quantisation results indicate substantial tolerance to degradation. On vehicle-mounted platforms the signal conflates object motion with platform motion, which we estimate at roughly 43% of the total on driving sequences. Whether ego-motion should count toward urgency is arguably an application question rather than a measurement one, since a rapidly moving camera does become stale faster, but it does mean that driving-only benchmarks partly measure vehicle speed.